sci-libs/faiss
Library for efficient similarity search and clustering of dense vectors
Screenshots
View
Download
BrowseChangeLog
commit 06a8fffd889e58a72561fba2dbbbf4b23899a431
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sat May 9 12:34:38 2026 +0200
sci-libs/faiss: new package, add 1.14.1
Facebook AI Similarity Search — efficient nearest-neighbour and
similarity search over dense vectors. C++ library with optional
SWIG-generated Python bindings (USE=python; PyPI distribution name
"faiss-cpu", import name plain "faiss").
Stepping stone toward sci-ml/amd-gaia's deferred ui extra (which boots
faiss eagerly under gaia.ui.server) and standalone-useful for any RAG /
semantic-search pipeline against sci-ml/sentence-transformers.
CPU-only build (no FAISS_ENABLE_GPU / cuVS / SVS / ROCm — those are
their own packaging worlds and the gfx1150 host has no Faiss-supported
GPU path). FAISS_OPT_LEVEL is hierarchical: cpu_flags_x86_avx512f >
cpu_flags_x86_avx2 > generic; the Python loader.py picks the highest
variant available at runtime. Verified end-to-end on this Strix Point
host (avx512 selected, IndexFlatL2/IP/IVF queries return sensible
results).
Quirk worth a comment: libfaiss_python_callbacks.so is a pure-C++
helper that swigfaiss DT_NEEDs. Upstream PyPI wheels keep it inside
the package dir and rely on RPATH=$ORIGIN, but Gentoo's cmake.eclass
strips RPATH at install time, so the ebuild relocates it to
/usr/lib64/. Verified 2026-05-09 the SONAME has no libpython
dependency, so the relocation is safe across Python versions.
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sat May 9 12:34:38 2026 +0200
sci-libs/faiss: new package, add 1.14.1
Facebook AI Similarity Search — efficient nearest-neighbour and
similarity search over dense vectors. C++ library with optional
SWIG-generated Python bindings (USE=python; PyPI distribution name
"faiss-cpu", import name plain "faiss").
Stepping stone toward sci-ml/amd-gaia's deferred ui extra (which boots
faiss eagerly under gaia.ui.server) and standalone-useful for any RAG /
semantic-search pipeline against sci-ml/sentence-transformers.
CPU-only build (no FAISS_ENABLE_GPU / cuVS / SVS / ROCm — those are
their own packaging worlds and the gfx1150 host has no Faiss-supported
GPU path). FAISS_OPT_LEVEL is hierarchical: cpu_flags_x86_avx512f >
cpu_flags_x86_avx2 > generic; the Python loader.py picks the highest
variant available at runtime. Verified end-to-end on this Strix Point
host (avx512 selected, IndexFlatL2/IP/IVF queries return sensible
results).
Quirk worth a comment: libfaiss_python_callbacks.so is a pure-C++
helper that swigfaiss DT_NEEDs. Upstream PyPI wheels keep it inside
the package dir and rely on RPATH=$ORIGIN, but Gentoo's cmake.eclass
strips RPATH at install time, so the ebuild relocates it to
/usr/lib64/. Verified 2026-05-09 the SONAME has no libpython
dependency, so the relocation is safe across Python versions.